ComfyDeploy: How ComfyUI_Diffree works in ComfyUI?

Youu can using Diffree in ComfyUI

Diffree: Text-Guided Shape Free Object Inpainting with Diffusion Model

---

Update

2024/10/30

• 优化一些代码，但是还有4G左右的占用一直没有被清理，有空再搞吧。因为推理用的autocast("cuda")，所以有些较旧的cuda卡可能会OOM

1.Installation

In the ./ComfyUI /custom_node directory, run the following:

git clone https://github.com/smthemex/ComfyUI_Diffree.git

2.requirements

if K_diffusion... missing module... check “no_need_requirements.txt”,pip missing module. K_diffusion 需求的库文件也在no_need_requirements.txt 里，缺啥装啥

pip install -r requirements.txt

如果是便携包的，提示缺少K-diffusion需要在comfyUI/python_embeded目录下，打开CMD pip安装需求文件;
if using portable standalone build for Windows 'comfyUI, you need in comfyUI/python_embeded;

python -m pip install -r requirements.txt

或者 or: python pip install -r requirements.txt --target "you path/comfyUI/python_embeded/Lib/site-packages"

Based on the SD model, ComfyUI users basically do not need to install any requirement libraries。
基于sd的模型，comfyUI安装版的用户，基本上不用装任何需求库

If a module is missing, please open 'nou_need-requirements.txt'
如果缺失库，请打开nou_need_requirements.txt文件看你少了啥

3 Need model

3.1 base model，只有1个，only a model: diffree-step=000010999.ckpt

├── ComfyUI/models/checkpoints/
|      ├── diffree-step=000010999.ckpt

3.2 vae
模型内置，并不需要 you can try some one

4 using tips

--推荐使用512或者类似的尺寸，他们用的256图幅训练，推荐的是320（稍微小了，现在SDXL当道）;
--Recommend using a size of 512 or similar, they use 256 frames for training, and the recommended size is 320;

5 Example

Text-Guided Shape Free Object Inpainting 常规文生内绘图演示。

Citation

OpenGVLab/Diffree

@article{zhao2024diffree,
  title={Diffree: Text-Guided Shape Free Object Inpainting with Diffusion Model},
  author={Zhao, Lirui and Yang, Tianshuo and Shao, Wenqi and Zhang, Yuxin and Qiao, Yu and Luo, Ping and Zhang, Kaipeng and Ji, Rongrong},
  journal={arXiv preprint arXiv:2407.16982},
  year={2024}
}